Cost modeling for agent teams

Agent teams have three cost layers — model, tools, and human review. The cheapest layer to overlook is the most expensive one.

Token cost is loud and easy to track. Tool cost is mid-loud. Human review cost — a senior person reading every output — is silent and usually the largest line item.

Plan in three columns

Model spend per task — tokens × price × volume
Tool spend — APIs, search, image gen, voice
Human spend — reviewer minutes × loaded rate

If your model spend doubles but your reviewer minutes drop in half, you saved money. Optimising the wrong column is how good agent rollouts look bad on paper.